Simultaneous Enhancement and Super-Resolution of Underwater Imagery for\n Improved Visual Perception

摘要

In this paper, we introduce and tackle the simultaneous enhancement and\nsuper-resolution (SESR) problem for underwater robot vision and provide an\nefficient solution for near real-time applications. We present Deep SESR, a\nresidual-in-residual network-based generative model that can learn to restore\nperceptual image qualities at 2x, 3x, or 4x higher spatial resolution. We\nsupervise its training by formulating a multi-modal objective function that\naddresses the chrominance-specific underwater color degradation, lack of image\nsharpness, and loss in high-level feature representation. It is also supervised\nto learn salient foreground regions in the image, which in turn guides the\nnetwork to learn global contrast enhancement. We design an end-to-end training\npipeline to jointly learn the saliency prediction and SESR on a shared\nhierarchical feature space for fast inference. Moreover, we present UFO-120,\nthe first dataset to facilitate large-scale SESR learning; it contains over\n1500 training samples and a benchmark test set of 120 samples. By thorough\nexperimental evaluation on the UFO-120 and other standard datasets, we\ndemonstrate that Deep SESR outperforms the existing solutions for underwater\nimage enhancement and super-resolution. We also validate its generalization\nperformance on several test cases that include underwater images with diverse\nspectral and spatial degradation levels, and also terrestrial images with\nunseen natural objects. Lastly, we analyze its computational feasibility for\nsingle-board deployments and demonstrate its operational benefits for\nvisually-guided underwater robots. The model and dataset information will be\navailable at: https://github.com/xahidbuffon/Deep-SESR.\n